Improved Hypertension Risk Assessment with Photoplethysmographic Recordings Combining Deep Learning and Calibration.

Area of Science:

• Cardiovascular disease research
• Artificial intelligence in healthcare
• Biomedical signal processing

Background:

• Hypertension is a major global health risk factor for cardiovascular diseases.
• Early detection and management of hypertension are critical for mitigating health risks.
• Photoplethysmography (PPG) offers a non-invasive method for physiological monitoring.

Purpose of the Study:

• To evaluate deep learning (DL) models (GoogLeNet, ResNet-18, ResNet-50) for classifying normotensive (NTS) and hypertensive (HTS) individuals using PPG signals.
• To assess the impact of varying calibration time intervals on the accuracy of DL-based hypertension detection.
• To determine the optimal calibration frequency for reliable PPG-based blood pressure assessment.

Main Methods:

• Utilized three deep learning architectures: GoogLeNet, ResNet-18, and ResNet-50.
• Employed photoplethysmographic (PPG) recordings for feature extraction.
• Investigated classification performance across different calibration intervals: <1 h, 1-6 h, 6-24 h, and >24 h.

Main Results:

• All tested DL models achieved over 90% accuracy with closely spaced calibration (<1 h).
• ResNet-18 demonstrated superior performance with 93.32% accuracy, 84.09% sensitivity, 97.30% specificity, and 88.36% F1-score for intervals <1 h.
• Classification accuracy decreased with longer calibration intervals, dropping below 81% for intervals >6 h but remaining above 71% for intervals >24 h.

Conclusions:

• Deep learning models, particularly ResNet-18, show high potential for accurate hypertension detection using PPG signals.
• Frequent calibration (intervals <1 h) is essential for maintaining high classification accuracy in PPG-based hypertension monitoring.
• These findings support the development of non-invasive, real-time, and reliable blood pressure monitoring systems.